The Cultural Divide�

Klaus Desmet

SMU, NBER and CEPR

Romain Wacziarg

UCLA and NBER

January 2019

Abstract

We study cultural convergence and divergence in the United States over time. Using the General

Social Survey, we document the evolution of cultural divides between groups, de�ned according

to 11 identity cleavages (gender, religion, race, income, region, education. . . ). Between-group

heterogeneity is small: the United States is very pluralistic, but this is primarily due to within-group

heterogeneity. On average, between-group heterogeneity fell from 1972 to the late 1990s, and grew

thereafter. We interpret these �ndings using a model of cultural change where intergenerational

transmission and forces of social in�uence determine the distribution of cultural traits in society.

�Desmet: Department of Economics and Cox School of Business, Southern Methodist University, 3300 Dyer, Dallas,

TX 75205, kdesmet@smu:edu; Wacziarg: UCLA Anderson School of Management, 110 Westwood Plaza, Los Angeles CA

90095, wacziarg@ucla:edu. We thank Omer Ali for outstanding research assistance. We also thank Alberto Alesina,

Raquel Fernández, Kai Gehring, Paola Giuliano, Ricardo Perez-Truglia, Fabio Schiantarelli, Jesse Shapiro, Francesco

Trebbi and seminar participants at numerous universities for useful comments.

1 Introduction

Many scholars and commentators have argued that the United States faces a growing cultural divide

along lines of race, geography, gender, age, income and other dimensions. These growing disagree-

ments go hand in hand with a fraying social fabric, growing dysfunction in the political arena, and

the disintegration of social capital. Others have argued that the greater availability of information

and exchange facilitated by travel and exposure to di¤erent cultures have brought about cultural con-

vergence, so that cultural heterogeneity between groups is becoming smaller as cultural traits di¤use

throughout society.1 Which view is correct?

In this paper, we conduct a systematic quantitative study of cultural convergence and divergence

in the United States over time. We assess whether cultural values - or memes - have grown more or

less heterogeneous across groups de�ned according to 11 identity cleavages, among which are gender,

religion, ethnic origin, family income quintiles, geographic region and education levels.2 We use the

General Social Survey (GSS), a survey of norms, values and attitudes, spanning 1972 to 2016. We

consider a wide range of memes covering religious beliefs and practices, con�dence in institutions,

preferences over public policies, moral values and attitudes, measures of trust and life satisfaction,

and tolerance for alternative viewpoints and lifestyles.

We use two classes of measures of cultural heterogeneity. The �rst captures overall heterogeneity,

describing, for each meme, the likelihood that two randomly chosen individuals surveyed in the GSS

will have a di¤erent cultural trait. The second is a measure of heterogeneity between groups, capturing

the degree of �xation of memetic traits onto group identity. A high degree of �xation indicates that

memes are highly group-speci�c, while a low degree of �xation indicates that the distribution of memes

within each group closely resembles that in society overall. Rising �xation, in this context, would be

associated with a growing cultural divide between groups.

We �nd that the overall degree of cultural heterogeneity in the United States is remarkably stable

when averaging heterogeneity across all available memes. We �nd evidence that average cultural

heterogeneity fell slightly between 1972 and 1993, and rose slightly thereafter. These average tendencies

mask interesting variation across questions. For some questions, such as several questions on sexual

behavior and public policies, there is growing social consensus. For others, such as questions on gun

laws and con�dence in some civic institutions, we �nd growing disagreements. Some of these dynamics

can be understood as transitions from one end of the belief spectrum to the other. For instance, on

the issue of marijuana legalization, attitudes have moved from generalized disagreement to majority

agreement, so heterogeneity rose and is now falling. Overall, we �nd some evidence of a systematic

tendency toward greater heterogeneity after 1993 when averaging over all available memes, yet on

many issues heterogeneity changes little.

1On the �rst view, prominent enunciations include Putnam (2000) and Murray (2013). Commentary along these lines

among pundits and journalists are too numerous to list. The second view is more closely associated with modernization

theory - see for instance Inglehart (1997) and Ritzer (2011), pointing, respectively, to rising incomes and globalization

as powerful forces for cultural homogenization.2The term "meme" was coined by Richard Dawkins (1989) to describe a cultural trait, much like a gene is a genetic

trait. A meme can take on several variants, for instance the meme "belief in God" could take on variants "yes" or "no".

A person�s culture, in our terminology, is simply that person�s vector of memetic traits.

1

Across all identity cleavages, the level of between-group heterogeneity is extremely small: we �nd

that most heterogeneity in cultural traits occurs within, not between, groups. The United States is

an extremely pluralistic country in terms of cultural attitudes and values, but this diversity is not

primarily the result of cultural divides between groups.

The time path of these cultural divides between groups displays interesting patterns. We �nd

evidence of falling between-group heterogeneity from 1972 to the late 1990s, and growing divides

thereafter, but only for some cleavages and only for some memes. Average heterogeneity has risen

across religious identities since the mid-1990s, and across education levels, income quintiles, ethnic

groups and racial groups since the early 2000s. The same trend is particularly pronounced across

groups de�ned by political party self-identi�cation. The cultural divide across party self-identi�cation

started to gradually increase at the end of the 1990s, and rose sharply in the �rst half of the 2000s. Of

course, this may re�ect the ability of individuals to more easily self-identify with a party that closely

matches their cultural beliefs (sorting) rather than cultural change within groups prede�ned by party

identi�cation.3 In many cases, the most recent levels of between-group heterogeneity do not surpasslevels reached in the early 1970s. We also �nd stable or falling cultural heterogeneity across di¤erent

regions of the United States, across urban categories, across age groups, and across genders. Some

of these �ndings come as a surprise in light of the public pronouncements concerning growing divides

across some of these identity cleavages.

How can we interpret these results in light of the popular commentary on the fraying social and

cultural fabric of the United States? We hypothesize that several forces are at play, and may operate

di¤erently depending on speci�c memes and speci�c identity cleavages. To understand these forces,

one needs to form a picture of how memes change over time. To do so, we propose a model of cultural

change. In our model, three forces explain the distribution of memes across and within groups and its

dynamics: intergenerational transmission, social conformism and the emergence of cultural innovations.

First, an individual�s vector of memes originates, with variation, from intergenerational transmission.

Second, agents tend to conform to the majority memes of their own group. Third, innovations in

values (particularly values initially held by a minority) can occur and spread through social in�uence.

These three forces determine the dynamics of cultural change in the model.

Our model provides simple comparative statics to understand the dynamics of cultural change in

light of characteristics of memes and characteristics of identity cleavages. We use the model as a lens

through which we interpret our empirical results. A crucial distinction is whether social in�uence

occurs mostly within or across identity cleavages. This depends on the manner in which members

of society interact across and within groups. Our model helps characterize conditions under which

the emergence of new communication technologies reinforce within-group conformism and weaken

between-group interactions. This is important to understand the di¤erential dynamics of the cultural

divide across cleavages. It may help understand why, for instance, the cultural divide across Party ID

are going up while divisions between rural and urban areas are diminishing. This trend is reinforced by

the possible sorting of people with di¤erent memes into speci�c identities (such as party a¢ liations),

3For instance, the gradual realignment of Southern Democrats with the Republican Party over time may imply greater

�xation of political preferences on party identi�cation.

2

a possibility that we explicitly allow for in our model. We also discuss how cultural divides may

change di¤erently across question categories. For instance, the emergence of cultural innovations such

as greater social acceptance for gay marriage or marijuana legalization can lead to greater cultural

divides if adopted at di¤erent rates across identity cleavages.

Our paper is related to a growing literature on the evolution of cultural traits. Our terminology

and overall approach to culture borrow from the literature on cultural evolution (Cavalli-Sforza and

Feldman, 1981, Boyd and Richerson, 1985, Richerson and Boyd, 2004, Henrich, 2015, Bell, Richerson

and McElreath, 2009). Recent work by economists also tries to better understand the causes and

mechanics of cultural change. Salient examples in this tradition include Bisin and Verdier (2000),

Kuran and Sandholm (2008), Olivier, Thoenig and Verdier (2008), Fernández (2014) and Guiso, Her-

rera and Morelli (2016). Another literature, originating in political science and sociology, examines

cultural change arising from modernization and globalization (Inglehart, 1997, Inglehart and Baker,

2000, Norris and Inglehart, 2009, Ritzer, 2011). Our work is also linked to wide-ranging scholarship

on cultural change and persistence (Alesina and Giuliano, 2015, Giuliano and Nunn, 2017). A related

literature focuses on the features and behavior of immigrants to draw inferences on the persistence of

cultural traits across generations (Giuliano, 2007, Fernández and Fogli, 2006, Luttmer and Singhal,

2011, Giavazzi, Petrov and Schiantarelli, 2016).

Drawing on the aforementioned literature on the evolution of cultural traits, social scientists have

also studied heterogeneity in cultural traits, which is our main focus here. An important recent

contribution by Alesina, Tabellini and Trebbi (2017) studies cultural heterogeneity in Europe using

two waves of the World Values Survey. Like us, Alesina, Tabellini and Trebbi (2017) are interested

in characterizing cultural convergence or divergence. However, their focus is on the evolution of

cultural di¤erences between European countries, using heterogeneity between US states as a point of

comparison. Instead, we focus on the US, consider a wide range of eleven identity cleavages, use a

distinct measurement framework and interpret our �ndings through the lens of a model of cultural

evolution.

Bertrand and Kamenica (2018) apply a machine learning algorithm to a variety of survey data

in order to analyze how well someone�s culture or consumption behavior predicts their gender, race,

income, education and political ideology. Our approach to the cultural divide is di¤erent along a

number of dimensions. First, while Bertrand and Kamenica (2018) consider other sources of data on

culture and behavior besides the GSS �including surveys on consumer behavior and time use patterns

�we analyze the cultural divide along a broader set of identity cleavages (including urbanicity, ethnic

origin, region, age, religion and work status). Second, their measurement framework di¤ers from ours:

they use machine learning to quantify the extent to which a person�s identity can be correctly classi�ed

by knowing that person�s culture and behavior. In contrast, we calculate how predictive identity is for

cultural values. Third, we propose a conceptual framework to help us interpret why some divides have

deepened and others have not. Fourth, we emphasize that it is di¢ cult to understand the dynamics

of the cultural divide between groups without paying attention to the evolution of overall cultural

heterogeneity in society. For instance, when certain values become more acceptable in society at large,

they often di¤use at di¤erent rates in di¤erent groups, giving rise to a deeper divide. Finally, our

results di¤er. While they �nd that over the past half century, with the exception of political ideology,

3

the cultural divide has not greatly deepened, we �nd that along some cleavages or for some values, the

cultural divide has actually diminished, while on some others it has followed a U-shaped or increasing

pattern.

Our work is also related to research on cultural and political polarization in the United States

(DiMaggio, Evans and Bryson, 1996, McCarty, Poole and Rosenthal, 2006, Fiorina and Abrams, 2008,

Gentzkow, Shapiro and Taddy, 2016, Boxell, Gentzkow and Shapiro, 2017). Finally, the present study

shares its measurement approach with a recent literature on the measurement of cultural heterogeneity

at the individual-level rather than at the group level, using either genetic or memetic data (Ashraf

and Galor 2013, Desmet, Ortuño-Ortín and Wacziarg, 2017).

2 Measurement and Data

2.1 Measurement Approach

To capture cultural heterogeneity and the cultural divide between identity cleavages, we start from

the measurement framework in Desmet, Ortuño-Ortín and Wacziarg (2017). Consider c = 1; :::; C

identity cleavages that each consist of groups kc = 1; :::;Kc. Consider also m = 1; :::;M memes that

each can take on values im = 1; :::; Im. For instance, c could be gender (kc = male, female) and m

could be belief in God (im = yes, no). We denote by sim the share of the total population that holds

variant im of meme m, and by skc the share of group kc in the total population. We denote by simkc

the share of group kc (de�ned over cleavage c) that holds variant im of meme m. For instance, this

could be the share of males that believe in God.

Overall heterogeneity is simply memetic fractionalization over the whole population. For meme m:

CFm = 1�ImXim=1

(sim)2

Averaging over memes, we get average memetic fractionalization - the probability that two randomly

chosen individuals from the entire sample hold a di¤erent variant of a randomly drawn memetic trait:

CF =1

M

MXm=1

CFm = 1� 1

M

MXm=1

ImXim=1

(sim)2

CF is a measure of memetic heterogeneity in the entire population, regardless of identity cleavages.

To derive a measure of the cultural divide between groups, we calculate FST measures of memetic

�xation. Heuristically, FST captures the share of heterogeneity that occurs between groups de�ned

by identity cleavages (Wright, 1949; Cavalli-Sforza et al., 1994; Desmet, Ortuño-Ortín and Wacziarg,

2017).4 We start by de�ning heterogeneity in meme m within group kc:

CFmkc = 1�ImXim=1

�simkc

�24Desmet, Ortuño-Ortín and Wacziarg (2017) also de�ned a �2 measure of between-group heterogeneity. This captures

the information content of a person�s identity in terms of that person�s cultural values (Cover and Thomas, 2006). Hence

this index will take on high values when cultural traits are very group-speci�c. In practice, FST and �2 are very highly

correlated, so it matters little which one we use. Due to easier computation we focus on FST .

4

Taking the weighted average over groups for a given identity cleavage c, we obtain the average within-

group heterogeneity for meme m, CFmc :

CFmc =

KcXkc=1

skcCFmkc

Finally FST for meme m de�ned over cleavage c is simply the share of the total heterogeneity that is

not attributable to within-group heterogeneity:

(FST )mc = 1�

CFmcCFm

(FST )mc takes on values between 0 and 1. When (FST )mc = 0, group identity carries no information

concerning an individual�s cultural value. When (FST )mc = 1, knowing a person�s identity is equivalent

to knowing their value, i.e. the meme is perfectly �xated on groups.5 As was the case for CFm, (FST )mccan be averaged over all memes m to obtain the expected cultural divide between groups de�ned over

cleavage c.

2.2 Data

Selection of questions. We use survey data from the General Social Survey (GSS) from 1972 to

2016, from the 31 waves that have occurred so far. The universe of all GSS questions across all waves

includes 5; 895 �elds, but many of these questions were asked only once, either in special modules of

the GSS appearing only in a single wave, or as time-speci�c questions (e.g. about a given presidential

election). The �rst �lter that we apply is therefore to require that a given question be asked in at

least two di¤erent waves, in order to obtain some time-series variation. This leaves us with 2; 363

questions.

Among these, questions fall into various types. To capture a respondent�s vector of memes, we

need to consider the universe of questions that refer to values and attitudes. To this end, we classi�ed

each question into one of three types: 1) questions clearly about the respondent�s attitudes and values

(820 questions), 2) factual questions that can be considered to re�ect the values of the respondent, for

instance, "do you have a gun in your home?" or "how often do you attend religious services?" (272

questions) 3) Questions not related to the respondent�s values, including those that relate to facts

about other people (the respondent�s spouse, parents, etc.), the respondent�s education when younger,

as well as identity or demographic questions. For the purpose of determining the set of cultural memes

used in the analysis, we only retain questions of the �rst two types.6 This results in a set of 1; 092

questions.

Baseline set of questions. The frequency with which these 1; 092 questions were asked over time is

highly variable. Some were asked more or less continuously across all waves while others were asked for

only a small subset of waves. For our baseline exercise, we require as long a time series as possible over

5 (FST )mc = 1 can only happen when the number of identity groups Kc is at least as large as the number of possible

cultural values Im, and there is no within-group heterogeneity in values.6All of the 11 identity traits are drawn from answers to questions of the third type.

5

a common set of questions, to ensure the comparability of the measures of cultural diversity across

time. The questions that are asked regularly in the GSS are also more likely to re�ect important,

salient societal and cultural issues - trust, life satisfaction, attitudes toward fundamental civil rights

like freedom of speech, etc.

There is a trade-o¤: the higher the frequency over which the measures are computed, the smaller

the set of common questions across successive observations. To achieve balance between these con-

siderations, we group the survey data into either two-wave sets or �ve-year sets and keep questions

that were asked at least once in each grouping (Appendix Table A1 displays these groupings). This

amounts to keeping questions asked at least once every two waves, or at least once in any �ve-year

period. In the end, we are left with 76 memes when requiring questions to be asked every two waves,

and 96 memes when requiring questions to be asked at least once in each �ve-year interval. These

questions are listed in the Appendix Table A2. We use the 76 questions obtained from the two-wave

groupings as our baseline set, since it provides higher frequency for the heterogeneity measures, i.e. 16

groupings computed from 31 waves.7 We use the expanded set of questions obtained from the �ve-year

groupings for robustness checks presented in the Appendix (this gives 9 time periods).

Question entry and exit. The analysis of cultural heterogeneity over questions that enter or exit

the survey at a given point in time could also be interesting. Many of these �eeting questions are

asked only episodically in special GSS modules devoted to deeper investigations of topical subjects.

But some questions may also enter or exit the survey depending on the degree of social consensus. Of

particular concern is the exit of questions for which a social consensus has developed, and the entry

of questions that are characterized by emerging divides. Entry and exit of questions along those lines

cause opposing biases on CF (it is hard to form priors on the direction of the bias on FST ).

To address these issues, we conduct a systematic analysis of question entry and exit. In an extension

to our baseline exercise, we include questions that enter and exit in our measures of overall and cross-

group heterogeneity, to assess the e¤ect they have on the dynamics of the cultural divide. The analysis

proceeds in two ways. First, we calculate the heterogeneity indices over the full set of 1; 092 questions.

Of course, the indices are based on sets of questions that vary greatly through time, so this exercise

is the polar opposite of our baseline analysis based on a time-invariant set of questions. Second, we

focus more speci�cally on questions that appeared repeatedly in the survey and then were permanently

removed, and conversely questions that did not appear and then were consistently included.

To do so, we apply a simple algorithm: we divide the sample period into two subperiods (1972-1989,

i.e. 16 waves, and 1990-2016, i.e. 15 waves). Next, we identify questions asked at least �ve times in the

�rst subperiod and never in the second ("exit" questions). There are 21 such questions. For instance,

a question on whether birth control information should be available ("pill") is asked in �ve waves in

the 1970s and early 1980s, and is then permanently dropped. Conversely, we identify questions never

asked in the �rst subperiod, and asked at least �ve times in the second ("entry" questions). There are

60 such questions. For instance, a question about a¢ rmative action in hiring and promoting women

appears �rst in 1996 and is asked in almost every wave thereafter ("fehire").

7Of these 76 questions, 64 are unambiguously about values and attitudes, while 12 are factual questions that we

classi�ed as re�ecting the respondent�s values, such as those on gun ownership or church attendance.

6

We observe that in the universe of questions on cultural values that are asked at least twice, the

number of questions that enter or exit as de�ned above is rather limited (7:5% of the questions). Most

questions appear only episodically, without a systematic pattern of sustained entry or exit. Having

identi�ed the set of questions that persistently enter and exit the survey, we can examine if their

inclusion in our indices of average heterogeneity a¤ects the dynamics we describe. We do so below in

Section 3.2.

Question categories and types. Questions come in di¤erent categories and types. We rely on the

question categories provided by the GSS to classify questions. Broad categories include civil liberties,

current a¤airs, gender and marriage, politics, religion and spirituality. These are further divided into

�ner subcategories. For instance, gender and marriage includes questions on children and working,

on marriage, and on sex and sexual orientation. Questions are either binary or answered on a scale.

In our baseline set of 76 questions, 26 are binary (yes/no, agree/disagree) and 50 admit answers that

can be ordered on a scale. In 35 cases, the scale admits 3 answers, and in the remainder, 4 or more

possible answers.

Identity cleavages. We consider 11 identity cleavages to compute the FST indices. These cleavages

are characteristics of the respondents also observed in the GSS survey waves. They are labeled age,

education, ethnicity, family income, gender, party ID, race, region, religion, urbanicity and work

status. These cleavages admit anywhere from two values (gender) to nine values (region, ethnicity),

with the modal number of categories equal to �ve. Table A3 in the Appendix displays the cleavages

and corresponding categories.

3 The Evolution of Cultural Divides in the United States

3.1 The Dynamics of CF and FST

The evolution of overall heterogeneity. Figure 1 displays the time path of CF , averaged over

all 76 questions available, and the �rst columns of Table 1 shows the underlying numbers. We �nd

that average CF varies between 0:482 (in 1993) and 0:500 (1976). There is a U-shaped pattern over

the sample period: overall heterogeneity declined between the early 1970s and the mid-1990s and

grew back to its initial level by the end of the period.8 This average over all questions masks some

underlying heterogeneity. Panel A of Table 2 breaks down the dynamics by question. We �nd that

14:5% of the questions display a signi�cant U-shaped pattern (with the minimum reached some time

between 1980 and 2005). Heterogeneity is declining for 29% of the questions and increasing for 25% of

them. The rest is either hump-shaped or �at. This �nding of a substantial degree of heterogeneity in

8The overall variation can re�ect a substantial change in the underlying shares of respondents giving a speci�c answer

to a question. For instance, consider a binary question. With a CF of 0:5, response shares would be equally divided

between both possible answers. Then a change in CF to 0:482 represents a shift in answer shares of 9:5% (shares of 40:5%

to 59:5% for each possible answer). More generally, given the speci�c distribution of the number of possible answers

among our baseline set of 76 questions, the theoretical maximal average level of CF is 0:63. The United States appears

to be be quite culturally diverse overall, but there is room for that diversity to grow.

7

the dynamics of cultural diversity across questions will be echoed when discussing �xation measures,

highlighting the fact that generalizations about cultural diversity are hard to draw.

Table 3 characterizes the dynamics of CF by question category and subcategory. Overall there

is a lot of variation in the dynamics of CF across question categories. We tend to �nd U-shaped

or increasing paths for questions on crime, economic well-being and life satisfaction, and decreasing

heterogeneity on questions regarding free speech.

The evolution of cultural �xation by identity cleavage. For each of the 11 cleavages, Figure 2

displays the time path of FST , averaged over all questions (the underlying data is in Table 1). Figure

2 reveals an interesting ranking of cleavages by level of �xation, some of them surprising in light of

public commentary on the cultural divide. The biggest cultural divides are between groups de�ned

by educational attainment, family income quintiles and religion. The smallest divides are between

genders, races and urbanicity. But across all cleavages, the absolute level of �xation is very low, on

the order of 1 � 3%. The high level of cultural pluralism in the US, then, is not primarily due to

diversity between identity cleavages, but mostly due to diversity within identity categories.

These levels of cultural �xation change substantially through time. On average, one can discern

an overall U-shaped pattern, whereby cultural divides decreased between 1972 and the late 1990s, and

rose thereafter.9 Yet this masks very di¤erent patterns across cleavages. These are easiest to see in

Figure 3, which plots the dynamics of average FST cleavage by cleavage. For instance average FST for

Party ID is relatively �at through the mid-1990s, starts to gradually increase in the late 1990s, and

then accelerates in the 2000s, reaching its maximum in 2016. Of course, it is possible that people with

given Party IDs have grown culturally more distinct, or that people with distinct beliefs have sorted

more e¤ectively into di¤erent party IDs. A similar pattern is found for religion, and to a weaker extent

for ethnicity. Other cleavages display �atter or mild U-shaped patterns: family income, education and

race. Finally, some cleavages show declining levels of cultural �xation, though the decline typically

�attens at the end of the sample: age, urbanicity, region, and work status. Average FST for gender is

mildly hump-shaped around a very low level. The latter patterns are once again surprising in light of

many commentators�priors on rising divides across urban categories, genders, regions of the US and

employment status.

Dynamics of FST across questions. Table 2, Panel B classi�es the types of dynamics of FSTacross questions for each cleavage. The �rst observation is that, across all 11 cleavages, about 50% of

the questions display no clear direction over time: the dynamics are �at. For the remaining questions

that do display signi�cant patterns, we largely con�rm the dynamics of average FST displayed in

Figure 3. For instance, for 48:7% of the questions, FST based on Party ID displays a signi�cant U-

shaped pattern over the sample period, with an additional 6:6% of the questions displaying a strictly

increasing trend. Similarly for religion, FST is U-shaped for 34:2% of the questions, and increasing for

5:3% of them. Positive trends are weaker for race and ethnicity, with a combined share of U-shaped

and increasing patterns equal to 36:8% and 32:9%, respectively. For region, urbanicity and age, we

9A simple average of cultural �xation across the 11 identity cleavages reaches a minimum in 1997, and starts to

increase in 2001.

8

see high shares of declining FST indices (respectively 34:2%, 27:6% and 31:6%). Finally for gender,

we see a combined share of hump-shaped and declining FST dynamics equal to 29:0%.

Table 4 classi�es the types of dynamics within question categories and sub-categories, for all 11

cleavages. We rely on the nomenclature of questions provided by the GSS. This gives �ve broad

categories (civil liberties, current a¤airs, gender & marriage, politics, and religion & spirituality) that

are further divided into sub-categories. For instance, for civil liberties there are 23 questions, and 11

cleavages: when we state that 19:76% of the dynamics are U-shaped we mean that 50 out of 11 � 23series have U-shaped dynamics. We �nd again that a generally large share of the questions display �at

dynamics of FST . But interesting patterns emerge nonetheless. For instance, for free speech, a large

percentage of the question-cleavage categories (52:5%) display signi�cantly decreasing levels of FST .

These same questions, incidentally, tend to display a decreasing CF , indicating that between-group

diversity is decreasing faster than overall diversity. Another notable category is the set of questions

on crime, where we �nd on the contrary that �xation is either U-shaped or increasing in about 42:7%

of the cases. A similar pattern is found for questions on sex and sexual orientation, with a combined

share of U-shaped and increasing FST indices equal to 49:1%.

Analysis of the level of FST . Table 5 carries out a regression analysis of variation in the level of

FST . We pooled all of the FST measures across cleavages, questions and periods (with 76 questions,

16 periods and 11 cleavages, this gives us 13; 376 observations). Each panel reports results on each

of three sets of regressors: cleavage type, question category or subcategory, and time period (these

are all entered simultaneously). We largely con�rm previous observations. Looking at Panel A, we

replicate the ordering of FST magnitudes across cleavages. The average level of FST is elevated for

age, education, family income and religion, and is low for race, urbanicity and gender (the latter is the

smallest, and hence is the excluded category). In sum, the ranking of FST magnitudes across cleavages

is consistent with that displayed in Figure 2. Panel B analyzes the level of �xation by question category

(column 1) and subcategory (column 2), �nding that across all cleavages, FST tends to be high for

free speech questions, sex and sexual orientation, and religious a¢ liation and behaviors. FST tends to

be low for national spending, children and working, and con�dence and power (the cultural divide on

questions on marriage is the smallest of all, which is why it is our excluded category). Finally, Panel C

looks at time e¤ects by including a dummy for each of the 16 periods (excluding the one that starts in

1972, which is the excluded category). We �nd a U-shaped pattern reminiscent of the general pattern

displayed in Figure 2: cultural �xation across all questions and cleavages tends to fall until the late

1990s, and to rise in the 2000s (the minimum is reached for the 1996-1998 wave grouping). Of course

these average level di¤erences mask a lot heterogeneity in time and across cleavages, already discussed

previously.

3.2 Robustness to the Choice of Questions

Alternative frequency. Appendix Tables A4-A7 and Figures A1-A3 replicate our baseline analysis

with the set of questions that appear in the GSS at least once every �ve years. The frequency of

observations is correspondingly coarser (9 time periods instead of 16), but the number of questions is

expanded (96 rather than 76). We uncover dynamics that are unchanged compared to the baseline

9

exercise: CF displays a U-shaped pattern over time while the increase in FST in the later part of the

period is particularly pronounced for Party ID and religion, as before. We continue to �nd �at or

decreasing cultural divides between age groups, genders, regions, work status and urban categories.

For the remaining cleavages the pattern is U-shaped, with divides by the end of the sample period

mostly returning to the level of the early 1970s. The analysis of the dynamics of cultural divides across

question categories (Tables A6 and A7) reveals no new insight compared to the baseline. We conclude

that expanding the set of questions by reducing the frequency of observations leads to �ndings that

are very similar to the baseline.

Question entry and exit. Appendix Figures A4 and A5 display the dynamics of CF and FSTincorporating questions that are not asked uniformly throughout the sample period. The �rst panel

of each row replicates results using the baseline set of 76 questions asked at least once every two

waves. The second panel shows these series obtained from the most expansive set of questions (1; 092

questions asked at least in two waves of the GSS). The third panel displays the evolution of CF and

FST for the baseline set of 76 questions, augmented with 21 questions that permanently exited the

survey at some point, and 60 questions that were at �rst never asked, and then asked consistently.

Figure A4 shows that �ndings regarding CF are quite di¤erent across the �rst two panels: with

the expanded set of questions, the average level of CF is higher, indicating that questions asked only

episodically tend to be more divisive. The dynamics of CF are also di¤erent: in the second panel, the

series rises monotonically from the start of the sample period, going from about 0:5 to about 0:6.

In contrast, our �ndings for FST broadly con�rm the baseline results. We expected the FST series

to display more volatility than those based on a common set of questions, because the averages are

constructed on a constantly changing set of questions, most of which are only asked episodically.

However, this was not the case: comparing the �rst and second panels of each row of Figure A5, we

see almost identical levels and dynamic paths for FST across most cleavages. The only exception is for

gender where we see a more pronounced rise in FST early in the period (still from a very low level),

and a stabilization rather than a fall in more recent times.

The baseline set of 76 questions and the expansive set of 1; 092 questions represent polar opposite

choices along a spectrum. The third panels of Tables A4 and A5 represent a compromise between the

two extremes. Here, we �nd that the dynamics of both cultural heterogeneity and the cultural divides

are almost the same as in the baseline. Thus, our baseline results are not a¤ected by possible bias

stemming from the fact that exiting questions could be more consensual, and entering questions more

divisive.

In sum, a consideration of any question asked at least twice, and the inclusion of questions that

systematically enter or exit the survey, do not change the basic �ndings reported in Section 3.1

regarding the levels and changes in cultural divides over time (FST ). We do �nd a more pronounced

rise in overall cultural heterogeneity (CF ) using the most expansive set of questions, compared the

the baseline series based on 76 questions.

10

3.3 Alternative Approach Based on Regressions

As an alternative approach to assessing the dynamics of the cultural divide, we use a regression

approach. For each meme m at time t, we run a regression of the following form:

y = �+

11Xc=1

Kc�1Xkc=1

�kcDkc + "

where y is the response of an individual to meme m at time t , � is a constant, Dkc is a dummy variable

taking on a value of 1 if the individual is in cleavage category kc, and " is an error term. With 16

time periods and 76 memes, this means we are running 1; 216 regressions. We record the total R2

from each of these regressions, a measure of the informativeness of all identity cleavages together, in

terms of cultural memes. We then calculate the partial R2 due to each set of cleavage dummies. To

do so, we rerun the above regressions excluding the set of dummies for the cleavage of interest (this

is an additional 1; 216 regressions for each of 11 cleavages). We refer to the R2 from these regressions

as the restricted R2. For each meme i at time t, we then take the di¤erence between the total R2 and

the restricted R2, giving us the partial R2 for the corresponding cleavage. For each time t, we then

average the total and the partial R2 over all memes.

This approach is related to measuring �xation using FST . The greater the explanatory power of an

identity cleavage for cultural values, the higher the corresponding partial R2 in the above regression.

Similarly, the FST for that cleavage will tend to be relatively high. One advantage of the R2 approach

is that all identity cleavages are entered jointly, so we are controlling for the e¤ect of other cleavages

when assessing the explanatory power of a particular cleavage.

The results are presented in Table 6 and displayed graphically in Figures 4 and A6.10 The overall

R2, i.e. the joint explanatory power of all cleavages, displays a U-shaped pattern and is minimized

for the 1996-1998 wave grouping (Figure 4). The level of the R2 itself is modest, going from 15:5%

in 1972-1973 to 11:4% in 1996-1998 and back to 15:1% in 2016. The ability of cleavages overall to

explain answers to these 76 GSS questions therefore has increased starting in the early 2000s, indicating

growing cultural divides in the last decade and a half.

However, this average pattern masks interesting di¤erences cleavage by cleavage. These di¤erences

largely replicate those found for FST , con�rming that the average partial R2 re�ects a similar concept

of informativeness of cleavages for memes as does FST . In terms of the average levels of partial R2

and FST , there is a clear correspondence, with high values of both indicators for age, education,

family income, party ID, region and religion (compare the last rows of Tables 1 and Table 6). These

similarities in terms of average levels extend to the time path of the indicators cleavage by cleavage.

This is most easily seen by comparing Figure 3 and Figure A6: the dynamics of partial R2 are broadly

similar to those of FST , cleavage by cleavage. Overall, partial R2 values for Party ID are relatively

�at until the early 2000s, after which they increase rapidly, almost doubling in the span of 15 years.

We uncover a similar pattern for religion, with an acceleration starting slightly earlier, in the second

10Table A8 and Figures A7-A8 in the Appendix replicate these results using 5-year frequency data, expanded to 96

questions. The results are very similar to those described here for the baseline exercise using a 2-wave frequency and 76

questions.

11

half of the 1990s. We �nd a hump-shaped partial R2 for gender, and falling partial R2 for urbanicity

and region, as we did when capturing �xation with FST .

4 A Model of Cultural Change

In this section, we propose a model of cultural change. The model builds upon ideas from the cul-

tural evolution literature in both anthropology and, more recently, economics. Among the earliest

contributions to model culture in an explicitly evolutionary context are Boyd and Richerson (1985)

and Richerson and Boyd (2004, 2005). These authors proposed a range of evolutionary mechanisms

explaining the dynamic paths of cultural traits where cultural traits evolve through mutation and

selection, much like genes but at very di¤erent rates, partly because, unlike genes, cultural traits can

be transmitted horizontally.11 Bisin and Verdier (2000) study the intergenerational transmission of

norms in an explicitly economic model where parents rationally choose which traits to pass on to

their children, to derive the degree of cultural heterogeneity of a stationary population.12 Bernheim

(1994) models conformism, assuming that it arises from social in�uence: social status enters the utility

function, so there is a penalty for not conforming. Such conformism can lead to persistent customs

as well as temporary fads. Bikhchandani, Hirshleifer and Welch (1992) contains a theory of fads and

culture whereby certain values can originate from small shocks to preferences and spread through

local conformism, leading to informational cascades and cultural change. Kuran and Sandholm (2008)

compare the dynamics of cultural evolution in isolated and integrated societies, by analyzing the role

of intergroup versus intragroup socialization and coordination. The goal is to understand the con-

ditions under which cultural integration occurs, and circumstances under which societies can retain

their original cultures. We build on all these contributions, but emphasize the role of cultural diversity

between and within identity groups, since our purpose is to study how and why the resulting cultural

divide changes over time.

The aim of our conceptual framework is two-fold. First, we seek to understand the drivers of

di¤erent dynamic patterns of CF and FST . The speci�c sources of cultural change that we model

include intergenerational transmission, conformism, and cultural innovations. Some of these sources of

cultural change may lead to cultural convergence between groups, whereas others may lead to cultural

divergence, or more complex non-monotonic dynamics. Second, our model provides us with a lens

through which to interpret our empirical �ndings. Depending on characteristics of memes, of identity

cleavages, and of the extent of cross-group versus within-group cultural di¤usion, our model predicts

di¤erent dynamic patterns for CF and FST . We then discuss these predictions and their origins in

light of the speci�c empirical patterns identi�ed in Section 3.

11Genes and cultural traits can also coevolve. Henrich (2015) contains further explorations in a similar vein.12Doepke and Zilibotti (2008) also explicitly model parents�choices of values to impart to their children as a function of

economic incentives. Lazear (1999) models an individual�s choice to learn languages, gain familiarity with other cultures,

and assimilate culturally, again as a function of economic incentives to trade.

12

4.1 Setup

Consider a society with one identity cleavage (e.g., gender) and one cultural meme (e.g., belief in

God). The identity cleavage has two identity traits k and �k (e.g., male or female) and the culturalmeme can take two values i and �i. Time is discrete, t = 1; 2; ::: . Identity groups are of equal

size, and for now we assume that an individual cannot choose her trait. Each agent has one child, so

that each generation is as large as the previous one. Cultural values are imperfectly transmitted from

parent to child. As an agent socializes, she may change her cultural value in two situations. First,

if she was born with the minority value and is sensitive to conforming to his group�s majority value,

she may switch to the majority value. Second, we allow for the emergence of cultural innovations,

meaning that one of the values becomes more socially acceptable. If an agent has a taste for adopting

cultural innovations, she may switch to the value that has become more acceptable. Before stating an

agent�s decision problem, we describe in more detail the di¤erent determinants of his culture: vertical

transmission, pressure to conform and the adoption of cultural innovations.

Vertical transmission and innate values. There is imperfect vertical transmission between a

parent and a child. In particular, a share � of children inherits the value of their parent, and a share

(1 � �) is born with the other value. The parameter � therefore measures the intensity of verticaltransmission.13 We refer to the value an agent is born with as his innate value. In the absence of

conformism and innovation, the utility an agent derives from his innate value is normalized to one.

Pressure to conform. As an agent socializes, he may perceive a bene�t from conforming to the

majority value of his group. One bene�t from conformism may be that agents who frequently interact

gain from coordinating on the same value; another reason may be that some agents do not like to stand

out by being di¤erent from their group�s mainstream view. The bene�t from conforming increases

in the share of the own group that holds the majority value, but is heterogeneous across individuals.

In what follows, we denote by sik the share of group k that holds the majority view (and by s�ikthe share that holds the minority value, where obviously s�ik = 1 � sik). When born, an individualdraws a random variable p from a uniform distribution with support [0; 1=�p]. The utility he gets from

conforming to the majority value is then 1psik if he was born with the minority value and as

1+ p s

ik if

he was born with the majority value, where � 0 is a utility premium from having been born with

the majority value. A higher �p indicates a higher expected level of intragroup conformism in society

overall.

Our setup does not allow for inter-group conformism per se. However, when discussing comparative

statics on �p , we will argue that a weakening of within-group conformism (a lower �p ) can be interpreted

as a strengthening of between-group conformism.

The di¤usion of cultural innovations. We de�ne a cultural innovation as an existing value that

becomes socially more acceptable or fashionable. A cultural innovation is simply a label attached to

a given value that makes that value more attractive to hold. Some agents may �nd it attractive to

13We do not endogenize �, in contrast to the approach in the classic paper by Bisin and Verdier (2000), where the

intergenerational transmission of culture results from purposeful decisions by parents.

13

adopt this value, and switch from the non-innovating to the innovating value. Suppose that j is the

innovating value. For an agent of group k, the bene�t of holding the innovating value is increasing

in the share of agents of group k that hold this value, but is heterogeneous across agents. When a

cultural innovation occurs, each agent draws a random variable r from a uniform distribution with

support [0; 1=�rk]. This determines an agent�s utility from imitating the innovating value, 1rsjk. A

higher �rk indicates a higher expected level of sensitivity to imitating cultural innovations, i.e. a higher

sensitivity to fads, fashions or social trends.

Cultural innovations di¤use within groups, but may evolve di¤erently in the two groups if �rk and

�r�k are very di¤erent from each other. We discuss below situations under which �rk and �r�k may be

more or less similar to each other.

4.2 Decision problem

We now analyze an agent�s value choice at a given time t. Denote by i the value held by the majority

of the agent�s group and by j the value experiencing an innovation, where j could be equal or di¤erent

from i. An agent born with value x in group k, after drawing variables p and r, decides which value

x0 to adopt by maximizing the following discrete choice problem:

u(x; k) = maxx02fj;i;xg

�I(x); I(i)

1 + I(x)

psik; I(j)

1

rsjk

�(1)

where

I(x) =

8<:1 if x0 = x

0 otherwise

I(i) =

8<:1 if x0 = i

0 otherwise

I(j) =

8<:1 if x0 = j

0 otherwise

To give an example, consider someone born with the majority value in a society where there is a

cultural innovation to the minority value. If she holds on to her majority value, she will get a utility

equal to maxn1+ p s

ik; 1o, whereas if she switches to the innovating value she will get utility 1

rsjk.

Laws of motion. Since individuals do not always keep the value they are born with, we denote

by zik(t) the share of people of group k born in period t with innate value i and by sik(t) the share

of people of group k with value i after solving the discrete choice problem. Our assumption on the

imperfect vertical transmission of values between a parent and a child implies that

zik(t+ 1) = �sik(t) + (1� �)(1� sik(t)) = (2�� 1)sik(t) + (1� �) (2)

Of course if � = 1, vertical transmission is perfect so that zik(t+ 1) = sik(t).

To derive the laws of motion that determine cultural evolution, we solve the discrete choice problem

(1), assuming that the random draws of p and r are independent. We start by analyzing the case where

14

the innovation occurs to the minority value �i. Depending on their draws of p and r, agents of groupk born with value �i may want to switch to value i. Similarly, agents of group k born with valuei may prefer value �i. In Appendix B1, we show that the ex ante probability that an individual ofgroup k born in period t+ 1 with value �i prefers value i is �psik(t+ 1)� 1

2 �p�rksik(t+ 1)(1� sik(t+ 1)).

Because of the law of large numbers, this is the same as the share of agents of group k born with value

�i that switch to value i. As for individuals of group k born in period t + 1 with value i, the sharethat prefers to switch to value �i is �rk(1�sik(t+1))� 1

2 �p�rk(1+ )sik(t+1)(1�sik(t+1)). These results

yield the following law of motion for the share of the population holding value i when the innovation

occurs to value �i:

sik(t+ 1) = zik(t+ 1) + �psik(t+ 1)(1� zik(t+ 1))� �rk(1� sik(t+ 1))zik(t+ 1)

+1

2�rk �ps

ik(t+ 1)(1� sik(t+ 1))((1 + )zik(t+ 1)� (1� zik(t+ 1))) if j = �i (3)

Next we turn to the case where the innovation occurs to the majority value i. The share of

individuals of group k born in period t+1 with value �i who prefer to switch to value i can be shownto be �pski (t+1)+ �rks

ki (t+1)� �rk �p(ski (t+1))2. The law of motion for the share of people holding value

i when the innovation occurs to value i then becomes:

sik(t+ 1) = zik(t+ 1) + �psik(t+ 1)(1� zik(t+ 1)) + �rksik(t+ 1)(1� zik(t+ 1))

��rk �p(sik(t+ 1))2(1� zik(t+ 1)) if j = i (4)

The above two laws of motion are di¤erence equations that describe the evolution of the majority

value. Of course, the two laws of motion of the minority value are the complements of the laws of

motion of the majority value. The laws of motion of the other group �k can be written down byanalogy. Appendix B1 gives further details. Note that if no one is sensitive to cultural innovations

(i.e. �rk = 0), or if there is no conformism (�p = 0), then these di¤erence equations simplify considerably

and become linear.

Choice of identity trait. Until now we have assumed that agents cannot choose their identity

trait. Of course, for some identity cleavages (e.g., party ID) an individual can freely choose identity

trait k or �k. In that case, at a given time t, the discrete choice problem of an agent born with value

x becomes

u(x) = maxfu(x; k); u(x;�k)g (5)

where u(x; k) and u(x;�k) are the outcomes of maximization problem (1) for an agent who, respec-

tively, chooses identity trait k and �k. We postpone the discussion of the laws of motion under thisscenario until Proposition 3.

4.3 Patterns of Cultural Evolution

In this section, we analyze di¤erent patterns of cultural evolution generated by our model. In doing

so, we focus on the cases that are most relevant to our empirical analysis.

15

Conformism. We start by exploring a society with no di¤usion of cultural innovation and with no

choice of identity traits. We are interested in understanding how the steady-state value shares, and

hence CF and FST , depend on the intensity of vertical transmission and the intensity of conformism.

As we will now see, the results depend crucially on whether the majority value is the same across

groups or not.

Proposition 1: Conformism. Consider a society with no di¤usion of cultural innovations (i.e.�rk = 0). Then, in steady state:

1. The majority share in each group is weakly increasing in the strength of vertical transmission

(�) and conformism ( �p);

2. If the majority value is the same in both groups, FST is zero and CF is weakly decreasing in the

strength of vertical transmission and conformism;

3. If the majority value is di¤erent in both groups, FST is weakly increasing in the strength of

vertical transmission and conformism, and CF is maximized (and equal to 0:5).

Proof. See Appendix B2.

This proposition is intuitive. The steady-state share of the majority value is increasing in the

pressure to conform (�p) and in the strength of the intergenerational vertical transmission of values

(�). With stronger pressure to conform, individuals have a greater incentive to switch to the majority

value. As a result, the steady-state majority share becomes larger. With stronger intergenerational

transmission of values, the constraint on how high the majority share can become is weakened. Taken

together, there is less intragroup heterogeneity when �p is larger and/or � is larger.

By increasing the steady-state share of the majority value, larger values for �p and � reduce within-

group cultural fractionalization. If both groups conform to the same majority value, this also reduces

overall cultural fractionalization. Since, in that case, there are no di¤erences between groups, FST is

zero in steady state. If the two groups conform to di¤erent majority values, then a higher � and/or

a higher �p leave the society�s overall cultural fractionalization unchanged, because the two groups are

assumed to be of equal size. In this case, the cleavage between groups deepens, thus increasing FST .

How can we extend this discussion to a consideration of between-group conformism? Individuals

from one group may be sensitive to the majority value of the other group. Of course, the importance of

this force would depend on the importance of interactions between groups. For example, if the intensity

of interactions between groups declines, individuals become less sensitive to the majority view of the

other group. If the majority values di¤er across groups, then in our interpretation becoming less

sensitive to the other group is akin to becoming more sensitive to one�s own group. This translates

into an increase in �p, and hence a higher FST . If, on the other hand, the majority value is the

same across groups, then allowing for inter-group conformism does not a¤ect the steady-state cultural

divide, since FST remains zero.

Proposition 1 has a simple corollary which states that if an exogenous shock switches the majority

value of one of the groups, the cultural divide between groups will increase.

16

Corollary 1: Switching of Majority Values. Consider a society with no di¤usion of culturalinnovations. Starting o¤ in a steady state where both groups conform to the same majority value,

assume the value of the majority switches in one of the two groups. In that case, society converges to

a new steady state with higher FST and higher CF .

This result is immediate. If initially both groups have the same majority value, their steady-state

value shares are identical, so that FST is zero. Consider a shock that turns the majority value of one of

the groups into the minority value. Irrespective of the magnitude of this initial shock, the steady-state

value shares of that group will switch. For instance, if the two values had shares of 0:2� 0:8 in bothgroups, these now switch to 0:8� 0:2 in one of the two groups. As a result, the steady-state aggregatevalue shares are 1=2, so CF is maximized. Given that both groups now conform to di¤erent majority

values, there is a growing divide between groups, so FST increases. This result can be applied to a

situation where shifting circumstances disrupt the existing consensus enough to make the majority

view change in one of the groups.

Cultural innovations. We now turn to analyzing the di¤usion of cultural innovations, while still

assuming that individuals cannot choose their identity trait. We focus on a situation in which both

groups start o¤ holding the same majority value and where the innovation a¤ects the minority value.14

Proposition 2: Di¤usion of Cultural Innovations. Starting from a situation in which both

groups have the same majority value and the same majority share, suppose an innovation occurs to

the minority value.

1. If conformism is su¢ ciently weak and di¤usion is su¢ ciently strong, the majority value switches

in both groups. During the transition, CF exhibits a hump-shaped path.

2. If conformism is su¢ ciently strong and di¤usion is su¢ ciently weak, the majority value stays

the same in both groups. During the transition, CF increases.

Proof. See Appendix B2.

Once again, this proposition is intuitive. If di¤usion is strong, and hence �rk and �r�k are high,

individuals have a strong propensity to adopt innovations. Fads di¤use easily, and eventually take

over, becoming the new majority norm. As the original consensus breaks down, there is initially

growing disagreement between individuals. However, as the old majority norm is replaced by a new

majority norm, agreement between individuals once again increases. This translates into a hump-

shaped transition path for cultural fractionalization. If cultural di¤usion is weak in both groups, the

cultural innovation increases CF . In both cases, if the strength of di¤usion of a particular cultural

innovation di¤ers across groups, this will lead to a growing divide across groups since the steady state

shares of each value will be di¤erent across groups, and FST will rise.

14 In practice, for many memes, the majority value is the same across groups, so focusing on the case where both groups

have the same majority value is reasonable. Appendix B2 analyzes what happens if initially both groups hold di¤erent

majority values. As for the cultural innovation, the more interesting case is when it occurs to the minority value. If,

instead, it occurs to the majority value, then it simply reinforces the share of people holding the majority view.

17

How can this proposition inform our understanding of inter-group cultural di¤usion? In our model,

the innovation a¤ects the same value in both groups. However, the adoption pattern may be di¤erent

in the two groups, to the extent that �rk and �r�k are di¤erent. For instance, if �rk is low and �r�kis high, then group k will be much less sensitive to the innovation than group �k. If interactionsbetween groups are frequent and intense, the sensitivity to cultural innovations in the two groups is

likely to be more similar.15 Hence we can interpret di¤erences between �rk and �r�k as having e¤ects

on cultural diversity akin to those of interactions between groups. If �rk and �r�k are the same, cultural

innovations are adopted to the same degree in both groups, leaving FST unchanged. Correspondingly,

if the intensity of intragroup di¤usion is di¤erent across groups, a cultural innovation will lead to a

growing divide between groups. These insights are summarized in the following corollary.

Corollary 2: Di¤erences in Intra-group Di¤usion. Starting o¤ in a steady state where �rk and�r�k are di¤erent, if this di¤erence becomes smaller, then FST falls.

Choice of identity trait. We now let individuals choose their identity trait. Consider an individual

born with the minority value in her identity group. In addition to holding on to the minority value

in her group or adopting the majority value of her group, she now has one more option: she can also

switch identity groups. This may be an attractive option if she is a conformist and her value is held

by the majority in the other identity group. The following proposition summarizes this insight.

Proposition 3: Choice of Identity Trait. In a society with no di¤usion of cultural innovationswhere the majority value of one group is the minority value of the other, then as long as the majority

shares are smaller than one,

1. FST is larger if individuals can choose their identity trait than if individuals cannot choose their

identity trait;

2. The greater the degree of conformism, the larger the di¤erence in FST between a situation where

individuals can choose their identity trait and one where they cannot.

Proof. See Appendix B2.

This proposition says that the cultural divide between groups increases if individuals can freely

choose their identity trait. Moreover, the increase in the cultural divide is larger if within-group

conformism is stronger. The intuition for these two results is straightforward. Take an individual who

holds the minority value in the group she is born into. If it is costless to switch groups, then she would

rather change to the group where her innate value is held by the majority, as opposed to changing

her value. That is, if changing identity trait is free, then it is better to change identity trait than to

change value. This leads to sorting of values along identity traits, and hence to a rising cultural divide

between groups. The average payo¤ from sorting into the identity trait where one�s innate value is

held by the majority is especially high if within-group conformism is strong. Hence, the incentive

to sort on the majority value is greater in societies where people care a lot about conforming to the

group.

15By interactions we mean communication, contact and cooperative exchange between groups, not unlike the meaning

of "contact" in Intergroup Contact Theory in social psychology (Allport, 1954).

18

4.4 Interpretation of Empirical Findings

We now use our conceptual framework to interpret our empirical �ndings. In doing so, we are not

providing a test of our conceptual framework. Rather, we are using the model as a lens through which

to interpret the patterns observed in the data. We start by looking at how the dynamics of FSTdepend on the type of identity cleavage, and we then turn to analyzing the dynamics of both CF and

FST for di¤erent question categories.

4.4.1 Cultural Evolution across Cleavages

Our empirical analysis documents that on average FST displayed a U-shaped pattern between 1972

and 2016, declining from the early 1970s until the late 1990s, and increasing since the early 2000s.

Underlying this average pattern there is substantial heterogeneity across identity cleavages. Starting

in the late 1990s and accelerating in the mid-2000s, we see important increases in FST for Party ID

and for religion, and to a weaker extent for ethnicity. Other cleavages display a �atter or mildly

increasing pattern for FST since the late 1990s. These include family income, education and race.

Some cleavages exhibit no signi�cant change in FST in the later period: age, gender, urbanicity, region

and work status.

Here, we seek to explain these di¤erences in the dynamics of FST across cleavages. How do model

parameters relate to the evolution of the cultural divide? In the context of our conceptual framework,

the main parameters of interest are the level of �p and the di¤erence between �rk and �r�k. An increase

in �p can be interpreted as either a strengthening of intragroup conformism or a weakening of inter-

group conformism. According to Proposition 1, this would lead to an increase in FST . The intuition

is as follows: in a society where the majority values di¤er across groups, an increase in �p implies a

stronger tendency to conforming towards the own-group norm, and hence a weaker sensitivity towards

the views of the other group. In the case where the majority value is the same in both groups, an

increase in �p has no e¤ect on FST . Therefore, on average a higher �p implies a deepening divide, and

greater �xation of values on identity traits. An increase in the di¤erence between �rk and �r�k can be

interpreted as a weakening of the inter-group di¤usion of cultural innovations. According to Corollary

2, this would also lead to an increase in FST . The intuition is, once again, easy to understand. If there

is less inter-group di¤usion of innovations, then the adoption of innovations is more likely to occur

with di¤erent intensities across groups, leading to a deepening divide.

Within-group and between-group interactions. The environment in which individuals interact

with each other a¤ects the level of �p and the di¤erence between �rk and �r�k, i.e. whether social

in�uence occurs mostly within groups or also between groups. In this context, the rise of new forms of

communication in the late 1990s and early 2000s may have led to di¤erential changes in our model�s

main parameters depending upon the speci�c cleavage under consideration.

We start by discussing under which conditions these new technologies might have facilitated the

creation of echo chambers that favor within-group interactions relative to between-group interactions.

For modern technologies to have such echo chamber e¤ects, one of two conditions must hold. They

must allow individuals with a certain identity trait who previously lacked contact with others with

19

the same identity to deepen their interaction, or they must allow individuals with di¤erent identity

traits who previously had frequent contact to isolate themselves with people of their own identity.

These conditions do not apply to all identity cleavages. For example, in the case of the urban-rural

identity cleavage, individuals in urban areas already mainly interacted with other individuals in urban

areas before the advent of the modern technologies such as the internet. The same is true for group

identity based on the region where individuals reside. For both the rural-urban identity cleavage and

the regional identity cleavage, modern communication technologies would therefore not have much

scope to enhance group-speci�c echo chambers, simply because groups de�ned by these cleavages were

already living in echo chambers to start with.

Another example of an identity cleavage where we would not expect new technologies to have an

important impact is the gender cleavage. Individuals of a particular gender do not live isolated from

others of the same gender. For example, women do not need the internet to have frequent interaction

with other women. In that sense, the gender cleavage is similar to the urban-rural cleavage in that

it does not satisfy the �rst condition for modern media to have a group-speci�c echo chamber e¤ect.

Of course, one could argue that social media allows people of a certain gender to isolate themselves

more from those of a di¤erent gender. However, given how much people of di¤erent genders interact

with each other on a day to day basis, it is unlikely that the internet can make women and men more

isolated from each other. In that sense, the gender cleavage does not satisfy the second condition for

modern communication technology to have a group-speci�c echo chamber e¤ect. The same is true for

the age cleavage, but perhaps to a lesser extent because people of di¤erent ages do not necessarily live

together and interact to the same degree that men and women do.

For other cleavages, such as party ID, religion, family income, work status, education and race,

arguably the conditions above hold, at least to some extent. In these cases, people with di¤erent

identity traits are neither totally isolated from each other, nor interacting intensively. Hence, for these

identity cleavages, modern communication technologies have the potential to create group-speci�c echo

chambers, by facilitating interactions within groups. In our model, this corresponds to a higher �p or

a bigger di¤erence between �rk and �r�k, and thus a rising FST .

Summarizing, we can classify the 11 identity cleavages into two categories along the echo chamber

e¤ect dimension. Four identity cleavages have little scope for group-speci�c echo chamber e¤ects

(urbanicity, region, gender and age), whereas all the other identity cleavages have more scope for

group-speci�c echo chamber e¤ects. An increase in FST would only be prominent for the latter

identity cleavages for which modern communication technologies either allows individuals with the

same traits to connect or permits individuals with di¤erent traits to disconnect. Consistent with our

discussion, we should therefore not see a rise in FST for cleavages such as urbanicity, region, gender,

and possibly age. In contrast, there is potential for an increasing level of FST for all other identity

cleavages: party ID, religion, family income, work status, education, ethnicity and race.

Choice of identity traits. For identity cleavages with scope for an echo chamber e¤ect (party ID,

religion, family income, work status, education, ethnicity and race), Proposition 3 suggests that we

should expect the e¤ect to be particularly important for cleavages along which individuals can freely

choose their trait. For example, individuals can choose their party ID. The payo¤ from changing party

20

ID in order to align individual values with those of the majority is greater if within-group conformism

is stronger (i.e. when �p is higher). By introducing a complementarity between within-group echo

chambers and sorting, this increases the �xation of party ID on values. In contrast, individuals

can typically not choose their race. Although internet and social media make it easier for isolated

individuals of a certain race to interact with others of the same race, it does not increase the sorting

of races on particular values. Hence, �xation on race does not further increase.

An additional observation stems from the ability to directly sort into groups on the basis of cultural

values: the increasing alignment between values and traits such as party ID implies that the distinction

between identity traits and values could become more blurred. In that sense, some group-speci�c

echo chambers are not unlike value-speci�c echo chambers. For example, Republican-leaning media

increasingly coincide with media promoting conservative values, and vice versa.

If the emergence of new forms of media facilitate sorting on values rather than on identity traits,

then the cultural divide on some identity cleavages may actually fall. For instance, conservative

members of di¤erent racial groups may have an easier time �nding each other on social media. If

so, this might lead to a narrowing cultural divide between races. This observation suggests that the

dynamics of the cultural divide across di¤erent cleavages may need to be considered jointly. In the

example above, if individuals of di¤erent races are increasingly sorting on values, this may to some

extent be equivalent to increasingly sorting on party ID. In that case, the drop in the cultural divide

across races would coincide with a rise in the cultural divide across party ID.

Dynamics of FST across identity cleavages. In terms of how the internet, social media and

cable TV news are expected to a¤ect the cultural divide across di¤erent identity cleavages, the above

discussion suggests that two dimensions matter: the scope of the echo chambers e¤ect, and the ease

of sorting into identity trait. Figure 5 shows this graphically in a two-dimensional matrix. We can

distinguish between three categories of identity cleavages.

A �rst category consists of identity cleavages for which there is little scope for an echo chamber

e¤ect: age, gender, region and urbanicity. In the absence of an echo chamber e¤ect, being able to

choose one�s identity trait is irrelevant in the sense that modern media does not further reinforce

sorting. That is, the complementarity between echo chambers and sorting is inoperative when there

is no echo chamber e¤ect. For the identity traits in the left half of Figure 5 we would therefore expect

no increase in FST . A second category consists of identity cleavages with scope for an echo chamber

e¤ect, but identity traits cannot be freely chosen: ethnicity, race, and to a lesser extent, family income,

work status and education. For the identity cleavages in the bottom-right quadrant of Figure 5 we

would therefore expect the introduction of modern media to have a moderately positive e¤ect on FST .

A third class consists of identity cleavages with echo chamber e¤ects for which the complementarity

between echo chambers and sorting is at work: party ID, and to a lesser extent, religion. For the

identity cleavages in the top-right quadrant of Figure 5 our conceptual framework therefore predicts

an increase in FST following the introduction of modern media.

These theoretical predictions are largely consistent with the empirical patterns seen after the

introduction of modern media and communication technologies. Since the late 1990s, �xation is

21

mostly �at for age, gender, region and urbanicity; it is mildly increasing for ethnicity, race, income,

work status and education; and it strongly increasing for religion and party ID.

4.4.2 Cultural Evolution across Question Categories

There is substantial heterogeneity in cultural evolution, not just across identity cleavages, but also

across question categories and within question categories. In what follows we start by focusing on two

question categories that exhibit a fairly homogeneous pattern across questions.

Crime. The �rst of these categories are questions related to crime. In 69% of questions pertaining

to crime CF exhibits either a U-shaped or an increasing pattern over time. For the subset of crime

questions for which FST is not �at, 67% display a U-shaped or an increasing FST path. What might

account for the U-shaped pattern in CF and FST for many of the crime questions? One obvious

candidate is the evolution of the violent crime rate and the property crime rate, both of which peaked

in 1991. There are many explanations for the decline in crime rates since then. They include more

and better policing, mass incarceration, the end of the crack epidemic, the introduction of legalized

abortion, and the decline in lead exposure, among others.

To see how the rapid decline in crime rates might have changed people�s attitudes towards crime

issues, it is useful to focus on a particular example. Take, for instance, the question in the GSS that

asks respondents whether courts deal too harshly or not harshly enough with criminals. In 1991, of

those surveyed by the GSS, 4% answered courts were dealing too harshly with criminals, compared

to 79% who said courts were not harsh enough. By 2016, those numbers had changed to 18% and

55%, respectively. There are two ways of interpreting these numbers in light of the precipitous drop

in crime rates. If the driving force in the decline in crime is a harsher judicial system, this change in

policy may push more people to believe the courts are too harsh. Under this interpretation, people are

not changing their preferences about how harsh the courts should be, but given that the courts have

become harsher, fewer people now believe the courts are not harsh enough. As a result, we would see

CF increase. If, instead, the driving force in the decline in crime is unrelated to the judicial system,

then people may change their preferences about how harsh the courts should be given that crime rates

are lower. In our model we would view this as a cultural innovation that increases the minority view

that courts are too harsh. In other words, there is an innovation to the minority value. Through

the parameter �rk, this leads to a changing cultural consensus in the direction of a growing minority

believing that courts are too harsh. In that case, cultural heterogeneity increases, since the overall

consensus that courts are not harsh enough is waning. Hence, according to Proposition 2, we should

expect CF to increase, because of a cultural innovation to the minority value.

At the same time, the view on crime has become more divisive across identity groups. Going back

to the question on the harshness of courts, consider the changing racial divide. In 1991, there was a

broad consensus across racial groups: only 3% of whites and 12% of blacks answered that courts were

treating criminals too harshly. By 2016, these shares had increased to 16% and 38%, respectively. One

way of interpreting these facts is that whites have a lower �rk for this particular value than blacks. The

sensitivity of each group to the cultural innovation di¤ers, because di¤erent groups may be a¤ected

22

di¤erently by, say, the increase in mass incarceration. Consistent with Proposition 2, if �rk di¤ers from

�r�k, the divide between groups increases when an innovation occurs, and FST increases.

Another interesting question in the crime category relates to the legality of marijuana use. Between

1972 and the early 1990s there was a growing consensus that it should be illegal, reaching a maximum

of 83% in favor of keeping it illegal in 1990. Since then, the consensus has completely shifted, and by

2016 only 39% were still in favor of keeping marijuana illegal. As with the question on the harshness

of courts, this would lead to an increase in CF since the early 1990s. In terms of FST , here as well

the susceptibility to the innovation di¤ers across groups. For example, blacks were less in favor of

legalization than whites in 1990; this had switched by 2016.

These examples illustrate that when circumstances change, in a way that a¤ects di¤erent groups

di¤erently, the pre-existing consensus may weaken (showing up as increasing cultural heterogeneity)

and there may be growing divides across identity groups (showing up as growing �xation). Looking

ahead, whether in the long run the pre-existing consensus is replaced by a new consensus or whether

the new steady state is a lack of consensus will depend on the speci�c question. For example, in

the case of marijuana the growing majority in favor of legalization is such that in recent years CF

has started to decline, suggesting that a new consensus might be emerging. Indeed, when the old

consensus is replaced by a new consensus, Proposition 2 predicts a hump-shaped path for CF .

Free speech. The second category of questions where we see homogeneous patterns is the one related

to freedom of speech. For 78% of those questions CF exhibits a decreasing pattern over time. And for

the subset of free speech questions for which FST is not �at, 65% display a decreasing FST path. As an

example, consider the question whether an atheist should be allowed to make a speech against religion

in your community. In 1972, 62% of those surveyed answered positively; by 2016, this percentage had

increased to 80%. This points to a long-term growing consensus in favor of free speech, thus leading

to a falling CF over time. In general, this increasing agreement happened across all groups. As an

illustration, consider how the question on free speech for an atheist changed across the rural-urban

divide. In 1972, 80% of those living in locations of more than 1 million favored free speech for atheists,

compared to 58% of those living in locations of fewer than 10; 000. In 2016, those numbers were 80%

and 78%, respectively. Hence, for this particular question on free speech, the rural-urban divide all

but disappeared. As a result, in this case FST converged to a number very close to zero.

In the context of our model, this can be viewed as the di¤usion of a cultural value across groups.

The end of McCarthyism, the civil rights movement, and the increasing level of education might have

led to a renewed commitment to the First Amendment. Not all groups took this change on board

simultaneously, but eventually it di¤used to all groups. This led to a decrease in the di¤erence between

�rk and �r�k. According to Corollary 2, this should lead to a decrease in FST . This is an example of

cultural convergence. Why do some changing values di¤use across groups and others do not? One

reason is that the issue at stake may a¤ect di¤erent groups very di¤erently. For example, the harshness

of courts may a¤ect African Americans di¤erently from Whites, whereas the issue of free speech does

not have a strong racial element.

23

Same-sex relations. Within question categories, some speci�c questions exhibit strong dynamics

that are worth highlighting in the context of our model. For example, the percentage of people

answering that homosexual relations were always wrong peaked at the end of the eighties, with 78%

in 1987; by 2016, that �gure had gone down to 39%. The decline was especially rapid in the early

1990s. Between two consecutive GSS waves, 1991 and 1993, the percentage dropped by nearly 10

percentage points. This increasing tolerance towards same-sex relations translated into an increasing

CF . This is consistent with Proposition 2: as the original consensus disintegrates, we initially see

rising disagreement in society, and hence an increase in CF . This has happened across groups, but

not at the same rate. Compare locations below 10; 000 inhabitants to those above 1 million. In 1990,

the share answering homosexual relations were always wrong was 83% and 78%, respectively. These

�gures stood at 45% and 35% respectively in 2016. Hence, both saw a drop, but the drop was faster

in urban areas. In the context of our model, this is a cultural change going from one consensus to a

di¤erent consensus, but at di¤ering rates across groups. Thus, FST increases in the transition.

5 Conclusion

In this paper, we have conducted a systematic analysis of the evolution of cultural heterogeneity in the

United States. We sought to assess growing concerns about deepening cultural divides between groups

de�ned along a wide range of identity cleavages. We considered eleven such cleavages, such as race,

gender, income quintiles, educational attainment, etc. Using answers to questions on values, attitudes

and norms - cultural traits that we refer to as memes, in reference to Dawkins�(1989) terminology - we

characterized the time paths of cultural divides. The picture that emerges from this analysis is not one

of a generalized deepening of cultural divisions. First, the degree of between-group cultural speci�city

is very small, as between-goup variation represents between 0:6% (for gender) and 2:4% (education)

of total variation: most variation in memes is within groups. Second, we �nd, on average, a U-shaped

pattern for our FST measure of cultural �xation: on average cultural divisions tended to fall from the

early 1970s to the late 1990s, and to rise in the 2000s. In most cases, FST remains below its earlier

peaks. Third, the data does not justify a sweeping conclusion that there are deepening cultural divides.

The increase in the 2000s is driven largely by cleavages such as Party ID. Many commentators have

focused on the cultural divide across political lines, ignoring trends across other divides and ignoring

heterogeneity across memes. Our paper in contrast took a more systematic approach of looking at a

wide range of cleavages and memes. This broader approach does not warrant a pessimistic conclusion

that the United States is experiencing cultural disintegration. The data suggests a more quali�ed

conclusion that cultural divisions have grown only since the late 1990s, only for some cleavages and

only for some memes.

We also provided a theoretical interpretation for the heterogeneity in the dynamics of cultural

divides across cleavages and memes. In our model, agents are born with cultural traits inherited with

variation from their parents. Social in�uence then triggers potential changes in these inherited traits,

because agents conform to the majority of their own group and because they respond to cultural fads

and innovations: social in�uence is a major force explaining cultural change. The degree to which

cultural change is group speci�c determines the evolution of cultural divides between groups.

24

The model suggests that the manner in which agents access information and interact with each

other has important e¤ects on the evolution of cultural divisions. If the predominant mode of in-

teraction is between-groups, cultural change will occur in a similar manner across group identities,

keeping FST low. If instead most interactions are within-groups and information is group speci�c, it

becomes more likely that cultural �xation increases as a result of a cultural innovation. For instance,

new information technologies such as tailored cable TV channels and online social media can under

some circumstances increase the relative importance of within versus between-group social interac-

tions, by creating echo chambers. The dynamics of cultural divides also depend on characteristics of

the cultural traits under consideration. For instance, since the mid-1990s, there is an increasing view

that the justice system is too harsh on crime, but this change has occurred di¤erentially across races.

African-Americans are more likely to �nd the judicial system too harsh than Whites. In terms of

our model, this happens because the susceptibility of each group to this speci�c cultural innovation is

di¤erent, creating a growing divide.

Our work can be extended in several directions. First, we have provided a comprehensive analysis

of the dynamics of cultural divides across several identity cleavages, but di¤erences in these dynamics

warrant a closer analysis of the factors a¤ecting each cleavage. Second, for each cleavage, we have

considered all groups jointly, but this may mask interesting patterns for speci�c group pairs. For

instance, the average divide between all races may follow a certain time path, but the speci�c divide

between Hispanics and Whites may follow a di¤erent pattern. Third, we have also treated identity

cleavages separately but interactions may be relevant: while men and women may not have drifted

apart culturally, it is conceivable that African American women could have drifted apart from White

men. Our methodology can easily accommodate such extensions, as FST can be calculated for speci�c

pairs of identity groups, or for groups de�ned by the intersection of several traits.

Ultimately, we are interested in the evolution of cultural heterogeneity because of its potential

e¤ects on social cohesion, social capital and the ability of di¤erent groups to reach agreements on

public policy. In this paper, we have described the evolution of cultural divides, but the question of

their impact on political economy outcomes such as public goods provision, voting, inequality and

economic interactions between groups remains an important topic for future research.

References

Alesina, A. and Giuliano, P. (2015). "Culture and Institutions", Journal of Economic Literature,

53(4): 898-944.

Alesina, A., G. Tabellini and F. Trebbi (2017), "Is Europe an Optimal Political Area?", NBER Working

Paper #23325.

Allport, G. W. (1954). The Nature of Prejudice. Cambridge, MA: Perseus Books.

Ashraf, Q. and O. Galor (2013), "The Out of Africa Hypothesis, Human Genetic Diversity, and

Comparative Economic Development," American Economic Review, 103(1): 1-46.

Bell, A., P. J. Richerson and R. McElreath (2009), "Culture Rather than Genes Provides Greater

Scope for the Evolution of Large-Scale Human Prosociality," PNAS, 106(2): 17671-17674.

25

Bernheim, B. D. (1994), "A Theory of Conformity", Journal of Political Economy, 102(5): 841-876.

Bertrand, M. and E. Kamenica (2018), "Coming Apart? Cultural Distances in the United States over

Time", NBER Working Paper No. 24771, June.

Bikhchandani, S., D. Hirshleifer and I. Welch, (1992), "A Theory of Fads, Fashion, Custom, and

Cultural Change in Informational Cascades," Journal of Political Economy, 100(5): 992-1026.

Bisin, A. and T. Verdier, (2000), "Beyond the Melting Pot: Cultural Transmission, Marriage, and the

Evolution of Ethnic and Religious Traits", Quarterly Journal of Economics, 105: 955-988.

Boyd, R. and P. J. Richerson (1985), Culture and the Evolutionary Process. Chicago: University of

Chicago Press.

Boyd, R. and P. J. Richerson (2005), The Origin and Evolution of Cultures. Oxford: Oxford University

Press.

Boxell, L., M. Gentzkow and J. M. Shapiro (2017), "Greater Internet Use Is Not Associated With

Faster Growth in Political Polarization Among US Demographic Groups", PNAS, 114 (40): 10612-

10617.

Cavalli-Sforza, L. L. and M. W. Feldman (1981), Cultural Transmission and Evolution. Princeton:

Princeton University Press.

Cover, T. M., and J. A. Thomas (2006), Elements of Information Theory. Hoboken: John Wiley &

Sons.

Dawkins, R. (1989), The Sel�sh Gene (2 ed.), Oxford: Oxford University Press.

Desmet, K., I. Ortuño-Ortín and R. Wacziarg (2017), "Culture, Ethnicity and Diversity", American

Economic Review, vol. 107(9): 2479-2513.

DiMaggio, P., J. Evans and B. Bryson (1996), "Have Americans� Social Attitudes Become More

Polarized", American Journal of Sociology, 102(3): 690-755.

Doepke, M. and Zilibotti, F. (2008), "Occupational Choice and the Spirit of Capitalism", Quarterly

Journal of Economics, 123(2): 747�793.

Fernández, R. (2014), "Cultural Change as Learning: The Evolution of Female Labor Force Partici-

pation over a Century", American Economic Review, 103(1): 472-500.

Fernández, R. and A. Fogli (2006), "Fertility: the Role of Culture and Family Experience", Journal

of the European Economic Association, 4(2-3): 552-561.

Fiorina, M. P. and S. J. Abrams (2008), "Political Polarization in the American Public", Annual

Review of Political Science, 11: 563-588.

Gentzkow, M., J. Shapiro and M. Taddy (2016), "Measuring Polarization in High-Dimensional Data:

Method and Application to Congressional Speech", NBER Working Paper # 22423.

Guiso, L., Herrera, H. and Morelli, M. (2016). Cultural Di¤erences and Institutional Integration",

Journal of International Economics, 99: S97-S113.

Giavazzi, F., I. Petkov, F. Schiantarelli (2016), "Culture: Persistence and Evolution", NBER Working

Paper # 20174.

Giuliano, P. (2007), "Living Arrangements in Western Europe: Does Cultural Origin Matter?", Jour-

nal of the European Economic Association, 5(5): 927-952.

Henrich, J. (2015), The Secret of Our Success: How Culture Is Driving Human Evolution, Domesti-

cating Our Species, and Making Us Smarter, Princeton: Princeton University Press.

26

Inglehart, R. (1997). Modernization and Postmodernization: Cultural, Economic, and Political Change

in 43 Societies. Princeton University Press.

Inglehart, R. and W. E. Baker (2000), "Modernization, Cultural Change and the Persistence of Tra-

ditional Values", American Sociological Review, 65: 19-51.

Kuran, T. and Sandholm, W. H. (2008), "Cultural Integration and Its Discontents", Review of Eco-

nomic Studies, 75(1): 201-228.

Lazear, E. (1999), "Culture and Language", Journal of Political Economy, 107(S6): S95-S126.

Luttmer, E. F. P. and M. Singhal (2011). "Culture, Context, and the Taste for Redistribution",

American Economic Journal: Economic Policy, 3(1): 157-179.

Murray, C. (2013). Coming Apart: The State of White America, 1960-2010. New York, NY: Crown

Forum.

Norris, P. and Inglehart, R. (2009). Cosmopolitan Communications: Cultural Diversity in a Globalized

World. Cambridge: Cambridge University Press.

Olivier, J., Thoenig, M. and Verdier, T. (2008). "Globalization and the Dynamics of Cultural Identity",

Journal of International Economics, 76(2): 356-370.

Richerson, P. J. and R. Boyd, (2004), Not by Genes Alone: How Culture Transformed Human Evolu-

tion, Chicago: University of Chicago Press.

Ritzer, G. (2011), Globalization - The Essentials. London: Wiley-Blackwell.

Wright, S. (1949), "The Genetical Structure of Populations," Annals of Human Genetics, 15(1): 323-

354.

27

28

Tabl

e 1

– CF

by

Tim

e Pe

riod

and

F ST b

y Ti

me

Perio

d an

d Cl

eava

ge

(2-w

ave

grou

ping

)

Tim

e Pe

riod

CF

F ST

Age

F ST

Educ

. F S

T Et

hnic

F S

T Fa

mily

in

com

e

F ST

Gen

der

F ST

Part

y ID

F S

T Ra

ce

F ST

Regi

on

F ST

Relig

ion

F ST

Urb

an

F ST

Wor

k St

atus

19

72

0.49

5 0.

025

0.03

0 0.

015

0.02

0 0.

006

0.01

6 0.

012

0.01

9 0.

024

0.01

5 0.

019

1974

0.

494

0.02

5 0.

026

0.01

4 0.

017

0.00

6 0.

013

0.00

9 0.

020

0.02

2 0.

016

0.01

8 19

76

0.50

0 0.

024

0.02

6 0.

013

0.01

7 0.

006

0.01

1 0.

009

0.01

7 0.

019

0.01

2 0.

017

1979

0.

499

0.02

3 0.

028

0.01

4 0.

018

0.00

6 0.

011

0.01

0 0.

018

0.02

2 0.

013

0.01

7 19

82

0.49

7 0.

020

0.02

7 0.

017

0.02

0 0.

006

0.01

4 0.

015

0.01

8 0.

019

0.01

3 0.

018

1984

0.

496

0.02

2 0.

028

0.01

6 0.

018

0.00

7 0.

013

0.01

0 0.

016

0.02

0 0.

011

0.01

8 19

86

0.49

4 0.

020

0.02

5 0.

019

0.02

0 0.

007

0.01

2 0.

015

0.02

0 0.

018

0.01

1 0.

017

1988

0.

486

0.02

1 0.

027

0.01

3 0.

018

0.00

7 0.

010

0.00

8 0.

015

0.01

9 0.

010

0.01

8 19

90

0.48

4 0.

017

0.02

3 0.

013

0.01

6 0.

009

0.01

0 0.

009

0.01

8 0.

019

0.01

1 0.

018

1993

0.

482

0.01

7 0.

019

0.01

3 0.

015

0.00

6 0.

011

0.00

9 0.

014

0.01

8 0.

009

0.01

5 19

97

0.48

9 0.

015

0.01

9 0.

012

0.01

4 0.

006

0.01

3 0.

009

0.01

1 0.

020

0.00

9 0.

013

2001

0.

491

0.01

5 0.

018

0.01

3 0.

017

0.00

7 0.

013

0.01

1 0.

012

0.02

1 0.

008

0.01

3 20

05

0.49

5 0.

012

0.02

1 0.

016

0.01

8 0.

006

0.01

9 0.

015

0.01

1 0.

023

0.00

9 0.

011

2009

0.

499

0.01

4 0.

025

0.01

8 0.

019

0.00

6 0.

018

0.01

2 0.

012

0.02

4 0.

009

0.01

2 20

13

0.49

6 0.

013

0.02

4 0.

016

0.01

9 0.

005

0.02

1 0.

012

0.01

2 0.

025

0.00

8 0.

013

2016

0.

491

0.01

3 0.

022

0.01

7 0.

017

0.00

5 0.

023

0.01

3 0.

012

0.02

7 0.

010

0.01

4 Av

erag

e 0.

493

0.01

8 0.

024

0.01

5 0.

018

0.00

6 0.

014

0.01

1 0.

015

0.02

1 0.

011

0.01

6 Ti

me

perio

d re

fers

to 2

-wav

e gr

oupi

ngs.

So

for i

nsta

nce

1972

refe

rs to

poo

led

data

ove

r the

197

2 an

d 19

73 w

aves

of t

he G

SS.

29

Tabl

e 2-

Dyn

amic

s of C

F an

d F S

T, b

y Cl

eava

ge

(2-w

ave

grou

ping

, 197

2-20

16, 7

6 Q

uest

ions

)

U

-sha

ped

Hum

p-Sh

aped

In

crea

sing

De

crea

sing

Fl

at

Pane

l A: C

F CF

14

.47%

13

.16%

25

.00%

28

.95%

18

.42%

Pa

nel B

: FST

Ag

e 11

.84%

9.

21%

10

.53%

31

.58%

36

.84%

Ed

ucat

ion

15.7

9%

3.95

%

13.1

6%

21.0

5%

46.0

5%

Ethn

icity

14

.47%

3.

95%

18

.42%

10

.53%

52

.63%

Fa

mily

Inco

me

5.26

%

1.32

%

14.4

7%

22.3

7%

56.5

8%

Gend

er

6.58

%

17.1

1%

11.8

4%

11.8

4%

52.6

3%

Part

y ID

48

.68%

3.

95%

6.

58%

6.

58%

34

.21%

Ra

ce

14.4

7%

3.95

%

22.3

7%

14.4

7%

44.7

4%

Regi

on

11.8

4%

1.32

%

2.63

%

34.2

1%

50.0

0%

Relig

ion

34.2

1%

0.00

%

9.21

%

14.4

7%

42.1

1%

Urb

anic

ity

15.7

9%

2.63

%

5.26

%

27.6

3%

48.6

8%

Wor

k st

atus

7.

89%

11

.84%

7.

89%

18

.42%

53

.95%

Pa

nel B

Ave

rage

16

.99%

5.

38%

11

.12%

19

.38%

47

.13%

N

ote:

Thi

s Tab

le d

ispla

ys th

e fr

actio

n of

que

stio

ns, a

mon

g th

e 76

in o

ur b

asel

ine

sam

ple,

for w

hich

CF

or F

ST fo

llow

s the

type

s of d

ynam

ics l

isted

in th

e fir

st ro

w, i

.e. U

-sha

ped,

hum

p sh

aped

, inc

reas

ing,

de

crea

sing

or fl

at. T

o as

sess

thes

e dy

nam

ics,

we

regr

ess f

or e

ach

ques

tion

its C

F / F

ST o

n a

time

tren

d an

d its

squa

re. I

f bot

h th

e lin

ear a

nd q

uadr

atic

term

s are

stat

istic

ally

sign

ifica

nt a

t the

5%

leve

l, an

d th

e ve

rtex

of t

he fi

tted

qua

drat

ic c

urve

is b

etw

een

1980

and

200

5, w

e ch

arac

teriz

e th

e dy

nam

ics a

s eith

er

U-s

hape

d or

hum

p-sh

aped

. In

all o

ther

cas

es, w

e ru

n a

linea

r reg

ress

ion

of C

F / F

ST o

n a

time

tren

d, a

nd

clas

sify

the

dyna

mic

s as i

ncre

asin

g, d

ecre

asin

g or

flat

dep

endi

ng o

n w

heth

er th

e co

effic

ient

on

the

time

tren

d is

signi

fican

tly p

ositi

ve, s

igni

fican

tly n

egat

ive,

or i

nsig

nific

ant,

resp

ectiv

ely.

30

Tabl

e 3

- Dyn

amic

s of C

F, b

y Q

uest

ion

Cate

gory

and

Sub

-cat

egor

y (2

-wav

e gr

oupi

ng, 1

972-

2016

, 76

Que

stio

ns)

Que

stio

n Ca

tego

ry

Que

stio

n Su

b-ca

tego

ry

# of

qu

estio

ns

U-s

hape

d Hu

mp-

Shap

ed

Incr

easi

ng

Decr

easi

ng

Flat

Civi

l Lib

ertie

s

23

21.7

4%

17.3

9%

17.3

9%

43.4

8%

0.00

%

Cr

ime

13

38.4

6%

7.69

%

30.7

7%

23.0

8%

0.00

%

Di

ffere

nces

& D

iscrim

inat

ion

1 0.

00%

10

0.00

%

0.00

%

0.00

%

0.00

%

Fr

ee S

peec

h 9

0.00

%

22.2

2%

0.00

%

77.7

8%

0.00

%

Curr

ent A

ffai

rs

23

0.

00%

4.

35%

43

.48%

21

.74%

30

.43%

Econ

omic

Wel

l-Bei

ng

4 0.

00%

25

.00%

75

.00%

0.

00%

0.

00%

Nat

iona

l Spe

ndin

g 11

0.

00%

0.

00%

36

.36%

18

.18%

45

.45%

Soci

al Is

sues

8

0.00

%

0.00

%

37.5

0%

37.5

0%

25.0

0%

Gen

der &

Mar

riage

14

28.5

7%

21.4

3%

21.4

3%

21.4

3%

7.14

%

Ch

ildre

n &

Wor

king

2

50.0

0%

50.0

0%

0.00

%

0.00

%

0.00

%

Li

fe S

atisf

actio

n 6

33.3

3%

0.00

%

33.3

3%

16.6

7%

16.6

7%

M

arria

ge

1 0.

00%

10

0.00

%

0.00

%

0.00

%

0.00

%

Se

x &

Sex

ual O

rient

atio

n 5

20.0

0%

20.0

0%

20.0

0%

40.0

0%

0.00

%

Polit

ics

13

15

.38%

0.

00%

15

.38%

23

.08%

46

.15%

Conf

iden

ce &

Pow

er

12

16.6

7%

0.00

%

8.33

%

25.0

0%

50.0

0%

Po

litic

al B

elie

fs

1 0.

00%

0.

00%

10

0.00

%

0.00

%

0.00

%

Relig

ion

& S

pirit

ualit

y

3 0.

00%

66

.67%

0.

00%

33

.33%

0.

00%

Belie

fs

1 0.

00%

0.

00%

0.

00%

10

0.00

%

0.00

%

Re

ligio

us A

ffilia

tion

& B

ehav

iors

2

0.00

%

100.

00%

0.

00%

0.

00%

0.

00%

N

ote:

Thi

s Tab

le d

ispla

ys th

e ty

pes o

f dyn

amic

s of C

F fo

r diff

eren

t cat

egor

ies a

nd su

bcat

egor

ies o

f que

stio

ns, a

s def

ined

by

the

GSS.

The

ty

pes o

f dyn

amic

s are

list

ed in

the

first

row

, i.e

. U-s

hape

d, h

ump

shap

ed, i

ncre

asin

g, d

ecre

asin

g or

flat

. To

asse

ss th

ese

dyna

mic

s, w

e re

gres

s for

eac

h qu

estio

n its

CF

on a

tim

e tr

end

and

its sq

uare

. If b

oth

the

linea

r and

qua

drat

ic te

rms a

re st

atist

ical

ly si

gnifi

cant

at t

he 5

%

leve

l, an

d th

e ve

rtex

of t

he fi

tted

qua

drat

ic c

urve

is b

etw

een

1980

and

200

5, w

e ch

arac

teriz

e th

e dy

nam

ics a

s eith

er U

-sha

ped

or h

ump-

shap

ed. I

n al

l oth

er c

ases

, we

run

a lin

ear r

egre

ssio

n of

CF

on a

tim

e tr

end,

and

cla

ssify

the

dyna

mic

s as i

ncre

asin

g, d

ecre

asin

g or

flat

de

pend

ing

on w

heth

er th

e co

effic

ient

on

the

time

tren

d is

signi

fican

tly p

ositi

ve, s

igni

fican

tly n

egat

ive,

or i

nsig

nific

ant,

resp

ectiv

ely.

We

then

sum

mar

ize th

ese

dyna

mic

s by

aver

agin

g w

ithin

que

stio

n ca

tego

ries /

subc

ateg

orie

s.

31

Tabl

e 4

- Dyn

amic

s of F

ST, b

y Q

uest

ion

Cate

gory

and

Sub

-cat

egor

y, a

vera

ged

acro

ss 1

1 cl

eava

ges

(2-w

ave

grou

ping

, 197

2-20

16, 7

6 Q

uest

ions

)

Que

stio

n Ca

tego

ry

Que

stio

n Su

b-ca

tego

ry

# of

qu

estio

ns

U-s

hape

d Hu

mp-

Shap

ed

Incr

easi

ng

Decr

easi

ng

Flat

Civi

l Lib

ertie

s

23

19.7

6%

5.53

%

13.8

3%

29.6

4%

31.2

3%

Cr

ime

13

23.7

8%

6.29

%

18.8

8%

14.6

9%

36.3

6%

Di

ffere

nces

& D

iscrim

inat

ion

1 9.

09%

0.

00%

0.

00%

18

.18%

72

.73%

Free

Spe

ech

9 15

.15%

5.

05%

8.

08%

52

.53%

19

.19%

Cu

rren

t Aff

airs

23

14.2

3%

5.14

%

11.4

6%

13.0

4%

56.1

3%

Ec

onom

ic W

ell-B

eing

4

2.27

%

0.00

%

22.7

3%

13.6

4%

61.3

6%

N

atio

nal S

pend

ing

11

12.4

0%

9.92

%

11.5

7%

14.0

5%

52.0

7%

So

cial

Issu

es

8 22

.73%

1.

14%

5.

68%

11

.36%

59

.09%

G

ende

r & M

arria

ge

14

16

.23%

5.

19%

11

.04%

20

.13%

47

.40%

Child

ren

& W

orki

ng

2 13

.64%

9.

09%

0.

00%

13

.64%

63

.64%

Life

Sat

isfac

tion

6 4.

55%

1.

52%

13

.64%

28

.79%

51

.52%

Mar

riage

1

0.00

%

0.00

%

0.00

%

0.00

%

100.

00%

Sex

& S

exua

l Orie

ntat

ion

5 34

.55%

9.

09%

14

.55%

16

.36%

25

.45%

Po

litic

s

13

17.4

8%

5.59

%

5.59

%

9.09

%

62.2

4%

Co

nfid

ence

& P

ower

12

15

.15%

5.

30%

6.

06%

8.

33%

65

.15%

Polit

ical

Bel

iefs

1

45.4

5%

9.09

%

0.00

%

18.1

8%

27.2

7%

Relig

ion

& S

pirit

ualit

y

3 18

.18%

6.

06%

12

.12%

30

.30%

33

.33%

Belie

fs

1 27

.27%

0.

00%

9.

09%

18

.18%

45

.45%

Relig

ious

Affi

liatio

n &

Beh

avio

rs

2 13

.64%

9.

09%

13

.64%

36

.36%

27

.27%

N

ote:

Thi

s Tab

le d

ispla

ys th

e ty

pes o

f dyn

amic

s of F

ST fo

r diff

eren

t cat

egor

ies a

nd su

bcat

egor

ies o

f que

stio

ns, a

s def

ined

by

the

GSS.

The

ty

pes o

f dyn

amic

s are

list

ed in

the

first

row

, i.e

. U-s

hape

d, h

ump

shap

ed, i

ncre

asin

g, d

ecre

asin

g or

flat

. To

asse

ss th

ese

dyna

mic

s, w

e re

gres

s for

eac

h qu

estio

n its

FST

on

a tim

e tr

end

and

its sq

uare

. If b

oth

the

linea

r and

qua

drat

ic te

rms a

re st

atist

ical

ly si

gnifi

cant

at t

he 5

%

leve

l, an

d th

e ve

rtex

of t

he fi

tted

qua

drat

ic c

urve

is b

etw

een

1980

and

200

5, w

e ch

arac

teriz

e th

e dy

nam

ics a

s eith

er U

-sha

ped

or h

ump-

shap

ed. I

n al

l oth

er c

ases

, we

run

a lin

ear r

egre

ssio

n of

FST

on

a tim

e tr

end,

and

cla

ssify

the

dyna

mic

s as i

ncre

asin

g, d

ecre

asin

g or

flat

de

pend

ing

on w

heth

er th

e co

effic

ient

on

the

time

tren

d is

signi

fican

tly p

ositi

ve, s

igni

fican

tly n

egat

ive,

or i

nsig

nific

ant,

resp

ectiv

ely.

We

then

sum

mar

ize th

ese

dyna

mic

s by

aver

agin

g w

ithin

que

stio

n ca

tego

ries /

subc

ateg

orie

s and

acr

oss a

ll 11

cle

avag

es.

32

Table 5 – Regression analysis of the Level of FST, by cleavage type, by question category and by subcategory, and by time period.

Categories Sub-categories Panel A - Cleavages Age 1.194 (13.56)*** 1.194 (13.27)*** Education 1.770 (20.10)*** 1.770 (19.66)*** Ethnicity 0.851 (9.67)*** 0.851 (9.45)*** Family income 1.127 (12.80)*** 1.127 (12.52)*** Gender (excluded) (excluded) Party ID 0.775 (8.80)*** 0.775 (8.61)*** Race 0.475 (5.39)*** 0.475 (5.28)*** Region 0.884 (10.04)*** 0.884 (9.82)*** Religion 1.477 (16.77)*** 1.477 (16.40)*** Urbanicity 0.435 (4.94)*** 0.435 (4.84)*** Work status 0.908 (10.31)*** 0.908 (10.08)*** Panel B - Categories and sub-categories Civil liberties 0.801 (14.11)*** - Crime 1.163 (6.85)*** - Differences and discrimination 1.424 (6.15)*** - Free speech 2.488 (14.42)*** Current affairs -0.206 (3.63)*** - Economic well being 0.986 (5.39)*** - National spending 0.552 (3.23)*** - Social issues 0.721 (4.15)*** Gender and marriage (excluded) - Marriage (excluded) - Children and working 0.214 (1.07) - Life satisfaction 0.815 (4.61)*** - Sex and sexual orientation 1.434 (8.00)*** Politics -0.756 (11.74)*** - Confidence and power 0.084 (0.49) - Political beliefs 0.761 (3.29)*** Religion and spirituality 1.288 (12.10)*** - Beliefs 0.826 (3.57)*** - Religious affiliation and behaviors 2.858 (14.26)***

(Continued)

33

Categories Sub-categories Panel C - Time Dummies 1972 (excluded) (excluded) 1974 -0.139 (1.31) -0.139 (1.28) 1976 -0.283 (2.67)*** -0.283 (2.61)*** 1979 -0.192 (1.81)* -0.192 (1.77)* 1982 -0.133 (1.25) -0.133 (1.23) 1984 -0.200 (1.88)* -0.200 (1.84)* 1986 -0.155 (1.46) -0.155 (1.43) 1988 -0.323 (3.04)*** -0.323 (2.97)*** 1990 -0.347 (3.26)*** -0.347 (3.19)*** 1993 -0.507 (4.77)*** -0.507 (4.67)*** 1997 -0.544 (5.12)*** -0.544 (5.01)*** 2001 -0.479 (4.51)*** -0.479 (4.41)*** 2005 -0.361 (3.40)*** -0.361 (3.32)*** 2009 -0.289 (2.72)*** -0.289 (2.66)*** 2013 -0.290 (2.73)*** -0.290 (2.67)*** 2016 -0.252 (2.37)** -0.252 (2.32)** Intercept -0.064 (0.34) 0.829 (7.83)*** R2 0.15 0.11 Number of observations 13,376 13,376

* p<0.1; ** p<0.05; *** p<0.01; t-statistics in parentheses; FST multiplied by 100 to improve readability. Time dummies refers to 2-wave groupings. So for instance 1972 refers to pooled data over the 1972 and 1973 waves of the GSS, and the dummy takes on a value of 1 if the FST measure is computed using these underlying waves, and zero otherwise.

34

Tabl

e 6

– O

vera

ll or

Par

tial R

2 , Ove

r Tim

e

Tim

e Pe

riod

R2 O

vera

ll Ag

e Ed

uc.

Ethn

ic

Fam

ily

Inco

me

Gen

der

Part

y ID

Ra

ce

Regi

on

Relig

ion

Urb

an

Wor

k St

atus

19

72

0.15

5 0.

013

0.01

3 0.

006

0.01

1 0.

003

0.01

0 0.

003

0.01

4 0.

014

0.00

7 0.

007

1974

0.

141

0.01

2 0.

010

0.00

7 0.

009

0.00

3 0.

008

0.00

2 0.

012

0.01

3 0.

007

0.00

6 19

76

0.13

2 0.

012

0.01

1 0.

005

0.00

9 0.

004

0.00

7 0.

003

0.01

1 0.

011

0.00

7 0.

006

1979

0.

136

0.01

2 0.

011

0.00

4 0.

010

0.00

4 0.

008

0.00

2 0.

011

0.01

3 0.

007

0.00

5 19

82

0.13

5 0.

011

0.01

1 0.

006

0.00

9 0.

004

0.00

8 0.

003

0.01

1 0.

011

0.00

6 0.

006

1984

0.

140

0.01

1 0.

014

0.00

8 0.

010

0.00

5 0.

011

0.00

2 0.

013

0.01

2 0.

005

0.00

7 19

86

0.13

6 0.

010

0.01

1 0.

008

0.01

0 0.

005

0.01

0 0.

002

0.01

1 0.

011

0.00

5 0.

006

1988

0.

134

0.01

0 0.

013

0.00

6 0.

011

0.00

5 0.

011

0.00

2 0.

011

0.01

0 0.

005

0.00

7 19

90

0.13

4 0.

010

0.01

2 0.

006

0.01

0 0.

006

0.01

0 0.

002

0.01

3 0.

010

0.00

6 0.

006

1993

0.

118

0.00

9 0.

010

0.00

5 0.

010

0.00

5 0.

010

0.00

2 0.

008

0.01

1 0.

005

0.00

5 19

97

0.11

4 0.

008

0.01

1 0.

004

0.00

8 0.

004

0.01

0 0.

002

0.00

6 0.

013

0.00

4 0.

004

2001

0.

122

0.00

7 0.

010

0.00

5 0.

010

0.00

5 0.

010

0.00

2 0.

008

0.01

4 0.

005

0.00

5 20

05

0.13

3 0.

007

0.00

9 0.

004

0.01

0 0.

004

0.01

7 0.

002

0.00

8 0.

014

0.00

5 0.

004

2009

0.

135

0.00

7 0.

011

0.00

6 0.

009

0.00

5 0.

014

0.00

2 0.

008

0.01

6 0.

005

0.00

6 20

13

0.13

6 0.

007

0.01

1 0.

005

0.00

9 0.

004

0.01

6 0.

002

0.00

7 0.

015

0.00

5 0.

005

2016

0.

151

0.00

8 0.

012

0.00

8 0.

010

0.00

4 0.

020

0.00

3 0.

008

0.01

9 0.

005

0.00

7 Av

erag

e 0.

135

0.01

0 0.

011

0.00

6 0.

010

0.00

4 0.

011

0.00

2 0.

010

0.01

3 0.

006

0.00

6 Ti

me

perio

d re

fers

to 2

-wav

e gr

oupi

ngs.

So

for i

nsta

nce

1972

refe

rs to

poo

led

data

ove

r the

197

2 an

d 19

73 w

aves

of t

he G

SS.

35

Figu

re 1

– C

F ov

er T

ime

(76

ques

tions

, 2-w

ave

grou

ping

)

36

Figu

re 2

– A

vera

ge C

ultu

ral F

ST o

ver T

ime

for 1

1 Cl

eava

ges (

76 q

uest

ions

)

37

Figu

re 3

– E

volu

tion

of F

ST fo

r Eac

h of

11

Clea

vage

s, o

ver T

ime

38

39

Figu

re 4

– T

otal

R2 a

cros

s All

11 C

leav

ages

, Ove

r Tim

e (7

6 qu

estio

ns, 2

-wav

e gr

oupi

ng)

40

Figure 5 – Classification of Identity Cleavages